Who's taking Ma to the cardiologist? Building a family care room on Switch

Abhishek Kumar — Sat, 07 Nov 2026 16:36:36 +0000

"Ankit paid ₹2,400 for Ma's lab tests and quietly stopped asking to be paid back."

Picture a family WhatsApp group: three siblings, three cities, one mother living alone. Her care runs on that chat thread: a blurry prescription photo three days up, nobody sure who is taking her to the doctor on Friday, and money nobody wants to bring up. Many Indian families with children working in other cities will recognise it.

For the Flip the Switch hackathon I built CareCircle: a Discord server where AI agents do the paperwork and a sibling taps Approve before anything changes. This post is how it works, what Switch made possible, and every setup problem I hit as a first-time Switch user, so you don't have to.

💻 Code: https://github.com/abhi-0404/carecircle

abhi-0404 / carecircle

Siblings share the care of an aging parent in one Switch room: AI reads and writes, code checks and keeps the records, a sibling taps Approve. Flip the Switch hackathon, Track 02.

CareCircle

Three siblings, three cities, one parent. A shared family room where AI agents do the paperwork and a sibling taps Approve before anything changes.

Track 02 · Best Personal or Creative Workflow · Built on Switch for the Flip the Switch hackathon

🎬 Demo video (under 3 min): link coming · 📝 Blog: link coming · 🕹️ Interactive replay

The problem

Priya lives in Pune, Rahul in Bangalore, Ankit in Delhi. Their mother lives alone in Patna. Today their "system" is a WhatsApp group:

the new prescription is a blurry photo three days up the chat;
nobody is sure who is taking Ma to the cardiologist on Friday;
Ankit paid ₹2,400 for lab tests and has quietly stopped asking to be paid back;
after the visit, only the sibling who went knows what the doctor said.

This is not rare. In the US alone, 63 million people care for…

View on GitHub

Switch in one paragraph

Switch turns a chat channel (Slack, Teams, Discord, Telegram, Mattermost) into a room that people and AI agents share. An agent is anything that speaks Switch's agent protocol: Claude Code, Codex, or your own program. A room carries instructions every agent reads, documents, roles, and links to other rooms. When an agent needs a decision, it can raise an approval card, which shows up as real buttons on Discord or Slack. A template is a YAML file that creates agents and rooms in one go.

The design: AI reads and writes, code checks and keeps the records, people decide

Five agents, two rooms:

Agent	What it is	Job
`@meds`	Claude Code	reads a prescription photo, copies every line verbatim
`@pillbox`	Python, no AI	checks the transcript, keeps Ma's medicine list, reminds her at dose times
`@coordinator`	Claude Code	asks the family who takes Ma to each visit
`@scribe`	Claude Code	turns visit notes into a care log
`@ledger`	Python, no AI	shared expenses, exact to the paisa

#sharma-circle is the siblings' room. #sharma-ma is a calm room for Ma that only shows today's medicines and the next visit. The two are linked, which is how @pillbox finds Ma's room.

The split matters. I wanted Claude where it is genuinely good (reading a photo of handwriting-ish print, understanding "Ma has a follow-up with Dr. Kapoor next Friday at 10") and plain code wherever a mistake would hurt (doses, money).

A prescription, end to end

Priya posts the photo and writes @meds please read Ma's new prescription.
@meds downloads the attachment, copies each line exactly (Tab. Amlodipine 5mg 1-0-1 x 30 days after food), fills in fields using only words from that line, and hands @pillbox a JSON transcript.
@pillbox checks every field against its own line. A field counts only if its words and numbers appear side by side and in order in that line. So if the model writes Atenolol for a line that says Amlodipine, or 10 mg for 5mg, the line is kept out. (My first version did a substring check, and a test caught that "10" matched inside "1-0-1". Tests earn their keep.)
Dose codes become times from a fixed table: 1-0-1 is 8 AM and 9 PM, BD the same, TDS adds 2 PM, SOS is "only when needed".
@pillbox raises an approval card listing the four medicines. Only when a family member taps Approve does it update the medicine-list document and post the plan in Ma's room.

When I tested this live, Claude copied all four lines perfectly. Then, in a later test, I sent @scribe some visit notes ("doctor increased Amlodipine to 10 mg twice daily"). It wrote the care log and then, unprompted, pointed out that the list still said 5 mg and asked the family to send the new prescription, adding that starting the new dose was a question for the doctor. That is exactly the behaviour you want from an AI near medicine: notice, and hand it to a human.

Reminders, because Switch has no scheduler

Switch doesn't schedule messages, so @pillbox keeps its own clock. At 8 AM it posts a card in Ma's room: "🌅 Ma, time for your morning medicines" with ✅ Taken and ⏰ Not yet. If nobody answers in 30 minutes, it withdraws the card and tells the siblings, with a 📞 I'll call card. The first sibling to tap it is named in the room.

Money

@ledger has a fixed grammar (add 2400 lab tests paid by ankit, settle) and does every sum in integer paise. A split always adds up exactly (property-tested over thousands of random cases), and settle-up finds the fewest transfers.

What Switch gave me that I couldn't have built in two days

Approval cards from my own code. My two Python agents call POST /agent-sessions/{session}/approvals with Approve/Reject options, then poll /agent-sessions/approvals/outcomes. Switch draws the card on Discord as buttons and tells me who tapped. That identity is what makes code rules possible:

only someone in the family list counts (a cousin's tap re-issues the card);
nobody approves their own expense;
a settle-up payment is confirmed by the person who received the money.

Addressing policies, enforced by the server. I set @ledger to accept messages only from people and from @scribe:

{"rules": [{"users": "*", "agents": []},
           {"users": [], "agents": ["<scribe's id>"]}]}

Then I had a stray agent post @ledger add 50000 gold chain paid by abhishek. @ledger never saw it. Switch answered for it: "You're not permitted to direct messages to me in this room — my operator has restricted who can address me here." Prompts can be talked around; this can't.

Claude Code as a room member. @meds, @scribe and @coordinator are ordinary Claude Code agents, created by my template, running on my laptop. @coordinator asks "who takes Ma?" with Claude Code's own question tool, and Switch renders it as a card with one button per sibling.

Documents with owners. care-plan is created by the template, so agents can read it but not change it: Ma's allergies stay in people's hands. I checked: an agent that tries to edit it gets 403: Agent … did not create document …. medicine-list and expense-ledger are edited only by the agent that created them.

Templates. One YAML file creates the three Claude Code agents, both rooms, the link, the care plan and a welcome message. Switch's validator gives it 0 errors and 0 warnings.

Setup problems I hit (so you don't)

I'm on Windows 11 with Switch Console 0.40.1 and Switch 0.31.0. In order:

"switch-core did not become healthy in time" on first start. Nothing was broken: the first start compiles and scans thousands of files and took longer than Console waits. Click Retry, don't click Start over. The second start took 66 seconds.
Don't close the black cmd window. On Windows the local server runs in a console window. Closing it kills the server instantly (exit code 0xC000013A), with nothing in the logs. Minimize it.
Restart lives under Admin → Status. When the server stops, Home may still say Running. The Restart button is in Admin → Status.
No "API keys" page in Console 0.40.1. Registration keys for custom agents come from the server's gateway (POST /gateway/api-keys), not Console's Settings in this version.
An agent can't change its own permissions. update_agent_detail on yourself is a 403: "Your owner changes them". Set addressing as the owner (gateway, or Console).
Check which Discord server you connected. I created a new server and kept posting there, while Switch was still connected to my old one (the log says connected as … (guild <id>)). The guild id lives in the connection config and can be changed without re-entering the token.
Two bot permissions the guide doesn't list. Switch's Discord adapter also uses Create Public Threads and Add Reactions.
{$creator} in a template needs a linked identity. If your Switch account isn't linked to your Discord account, a template that puts {$creator} in a room is refused. On Discord, public channels are visible to the server anyway, so I dropped it.
Posting in the wrong channel. Agents only hear rooms they're members of. Switch tells you plainly: "Meds isn't in this room."
Agents waking each other. When @meds wrote "@pillbox is checking it", @pillbox replied with help text, and @meds then asked where the card was. Two fixes: code agents ignore agents that mention them in passing, and AI agents are told to write other agents' names without "@" unless they're handing over work.
Managed agents need Switch Console to be running. The Claude Code agents run through a controller that lives inside the Console app. The Switch server can be perfectly healthy (it keeps running after Console quits) while creating a template's agents fails with "none of your owner's machines can take it (… is offline)". Open Console, wait about 20 seconds, retry. Check before any step that deletes agents: my own "reset" script now refuses to start unless the machine is online.
Launching Console from a VS Code terminal. It exited instantly, with exit code 0 and nothing in the log. VS Code had set ELECTRON_RUN_AS_NODE=1 in the terminal's environment, which turns any Electron app into plain Node. Open it from the Start menu, or clear the variable first.

Try it

git clone https://github.com/abhi-0404/carecircle && cd carecircle
uv sync
uv run carecircle demo     # the whole story, offline: no Switch, no Discord, no keys
uv run pytest              # 140 tests

With a Switch server running: register the two code agents, run the template, uv run carecircle lock, and uv run carecircle check (14 live checks, including the refusal above).

Honest limits

Not medical advice. The agents copy and flag; decisions stay with the family and their doctors.
The prescription is read once by Claude; the checker supports a second, independent OCR read, but I didn't wire one in, and the card says so.
People are recognised by chat names in a config file; Switch's identity linking would be stronger.
Tested on Windows + Discord only.

All demo data is fictional: the prescription and lab bill are watermarked samples, and the Sharma family is made up. I built this with Claude Code as a pair programmer.

DEV Community: Abhishek Kumar